C[omp]ute

From: "andrew cooke" <andrew@...>

Date: Fri, 11 May 2007 09:58:46 -0400 (CLT)

Recently there has been discussion on the Erlang mailing list about "monads".

Monads were popularised by Haskell - they are defined by type constructor
(something that "makes" monads) and two operations: unit and a binding
operation.  And they are used to impose a sequence on a set of operations.
 This is necessary because Haskell is lazy - the order in which functions
are evaluated is not the order in which they are defined in the program -
and so some kind of ordering is needed when doing things that involve
mutable state (for example, asking the user to enter data should come
before reading input!).

OK, but what has this got to do with Erlang?  Very little - Erlang is an
eager language and evaluates functions in a nice, easy-to-understand way. 
So there is no need for ordering.

Monads provide more than ordering - it it turns out that the "monad API"
is quite fundamental, and can be applied to many different scenarios.  So
it provides a unifying concept across lists, optional values, exceptions,
etc.

But, again, the Erlang discussions have nothing to do with this.

In fact, the Erlang discussions are not about the semantics of monads at
all.  The posters are purely concerned with the "syntactic sugar" that
Haskell added to support monads.  This lets you write code in an
"imperative style" - it gives the appearance of mutable values.

So what is really happening?

Partly this is a social phenomenon - it's a lot more acceptable to say "we
need monads" (which are a "cool" concept) than to admit that you want to
program using a style unsuited to the paradigm supported by the language.

But that's not the whole explanation.  This is clear from Haskell, where
the monad syntax is not used for "most" code.  Examples of elegant code,
which are abundant for Haskell, use Haskell's succinct and powerful
"fucntional" syntax.  It, in fact, a "newbie error" to write a program
entirely using the monad-specific "do syntax".

So it's clear it's possible write clear, elegant programs in a functional
style (even when the monad syntax is available).  Why don't people want to
do this in Erlang?

Part of the answer to that question is, again, social.  Erlang is growing
rapidly from a small user base.  So there is a lot of pressure from
incoming programmers, accustomed to procedural programming, who want what
they are accustomed to having.

But part of the answer is also found in Erlang's poor support for common
functional programming idioms.

Take the following line of code, which I wrote earlier today, as an example:

  NonBlankLines =
    filter(compose_1(fun util:not_/1, fun util:is_blank/1), Lines).

This filters a list of lines, discarding those which are not blank.  It
could just have well have been written as:

  NonBlankLines = filter(not is_blank, Lines).

In other words - Erlang *punishes* you when you write code in a functional
style.  It requires you to add a whole pile of pointless noise.  For
example, in the code above:

- compose_1 is necessary because Erlang does not have a syntax for
composing functions.  All I want to do is join two functions together, so
the first processes the output from the second.  This is a basic building
block in functional programming and could be supported by simply writing
the two fucntion names together ("not is_blank", as in the second
version).

- not_ is a utility function necessary because there is no way (that i
could fine) to refer to the inbuilt function "not" directly.  So "not_" is
defined as "not" just because inbuilt functions are, in some odd sense,
broken.

- namespace:name is necessary because even though i have imported these
functions the namespace is *still* necessary when referring to a function
directly.  There's no ambiguity - I could have used "is_blank" alone
(without "util:") if I just wanted to call the function.  This strange
restriction is only necessary when referring to the function's name.

- fun is necessary to flag functions.  Again, this is pointless - when I
call a function I don't need to specify what it is.  Erlang code
religiously follows a naming convention that separates functions and
variables (by case of the first letter) so this doesn't even require any
type analysis.

It's crazy.  The parser is brain-dead lazy.  No wonder people don't "get"
it and keep trying to turn Erlang into C or Java.

Andrew